Judgment, then control flow. A first Jev call reads the ticket, and code routes it to an agent. A second call sizes the task, and code gives that agent its LLM and reasoning depth. The console shows every real exchange.